Tangermeme 1.4.0 Adds Agentic Support For Genomic Sequence Models

GitHub: https://github.com/jmschrei/tangermeme Install: `pip install tangermeme` Skills: `tangermeme-install-skills` (bundled with the installation)

@AnthropicAI tangermeme is meant to be a foundation for both human- and agent-driven genomic analysis: modular primitives you (or, now, your agent) can compose without reinventing the basics. Always eager to get feedback or comments about what would be most useful for the community.

The last release cleaned up tangermeme by adding more documentation, expanding the unit tests (~400 -> ~1000), function input validation and warnings to guide agents, and type hinting. Things that both agents and humans would find helpful...

Here's a simple example. I want to inspect a trained Cherimoya model. It's not hard to create a simple Jupyter notebook but I'm lazy and want Claude to do it for me.

When I run this, it loads up the tangermeme skill...

A challenge I've had with LLM-generated code is that it is verbose and unverified. As a library of modular primitives for S2F modeling, tangermeme seemed like the perfect basis for more auditable code -- either as stand-alone scripts or in Jupyter notebooks.

This work has been partially supported by the AI for Science program from @AnthropicAI. It is just the first step of several that will be coming out soon, with the goal of fully automating all S2F work (scientific discovery, variant effect prediction, DNA design).

Here is an example from a current project where I wanted to time design algorithms using my new Cherimoya models. I didn't even mention tangermeme in the prompt, just the names of design algorithms, but it still loaded up the skill and plugged them in appropriately.

At the top it's loading up the relevant tangermeme functions (loading, prediction, ISM, plotting)...

Uses them correctly...

And created a plot of ISM attributions.

The user did not need to know about these functions, how to use them, or even that they existed. But the agent discovered them and used them rather than spending effort + tokens writing it's own implementations. And now there's barely anything to audit!

Like many people, over the past few months, I've transitioned from writing code all day to writing prompts + auditing code. These skills have been such a relief for me because I know the functions work as long as they're called correctly.

@AnthropicAI After writing this thread, I noticed that this skill has enabled Claude to use tangermeme functions automatically in the small scripts it writes to confirm things.